Research On Image Random Impulse Noise Removal Model Based On Deep Learning

In the process of image generation and transmission,noise is a common problem that significantly affects the quality and usability of the images.Therefore,noise removal is an important step in image processing,as it helps to improve the clarity and understanding of the image content,which is particularly important in the field of computer vision.This thesis aims to study the model for removing random impulse noise from images,with the following main work:(1)Image random impulse noise removal model based on weighted densely dilated convolutional network: Existing deep learning-based image denoising methods often do not fully utilize the feature information at different levels.Channel merging is typically done by directly concatenating feature maps along the channel dimension,without considering the importance of shallow and deep convolutional features.To address this issue,a weighted dense dilated convolutional connection network model WDDCNet is proposed for removing random impulse noise from images.Firstly,the model enriches the multi-scale feature information of shallow feature maps using a context information block based on dilated convolution,which carries many fine-grained details of the image.Then,a weighted dense dilated convolutional connection block is used,with multiple dilated convolutions that integrate information obtained from small receptive fields into larger ones.To prevent loss of image detail information due to network depth,dense connections are used to pass shallow convolutional features to the deep layers.During dense skip connections,weights are introduced for each feature channel to enhance important information and reduce the impact of irrelevant and unimportant information.Finally,by integrating the output feature maps of the weighted dense dilated convolutional connection block and background information block,the restored image is obtained.Experimental results show that compared to other methods,this network has better overall denoising performance,effectively removing random impulse noise from images and better preserving edge and texture details.(2)Image random impulse noise removal model combing convolution and SwinTransformer: Swin-Transformer has the property of long-range dependency,which limits its ability to obtain rich local information.In contrast,convolutional neural networks have the advantage of local perception,allowing them to capture spatially local image features,but with weak feature correlation in the region of feature extraction.To address this issue,a novel image random impulse noise removal network model CSTNet based on convolution SwinTransformer is proposed.The network is mainly composed of recursive residual groups,which adopt a hierarchical structure in the recursive residual groups,and different layers learn different features.The first two layers use Swin-Transformer and convolutional layers respectively,and the last layer combines the color noise image through quaternion convolutional layer.During information propagation,some of the learned features are passed from top to bottom to the next layer,enabling information transfer and interaction and learning richer features.The network first passes through a convolutional layer,then through recursive residual blocks,and finally residual learning is used to restore the image.Experimental data shows that the proposed model has excellent denoising ability.(3)Image random impulse noise removal model of improved weighted dense dilated convolutional network based on selective kernel feature fusion and Swin-Transformer:Compared with WDDCNet and CSTNet,CSTNet has advantages in parameters and denoising performance,so WDDCNet has room for improvement in terms of parameters and denoising performance.Therefore,an improved weighted dense dilated convolutional network image random impulse noise removal model based on selective convolution kernel feature fusion and Swin-Transformer is proposed.The model improves the context information block and weighted dense dilated convolutional connection block,replacing feature fusion with attentionbased selective receptive field feature fusion,achieving adaptive receptive field and more accurate feature extraction at different scales.The recursive residual groups from CSTNet is also used to enhance the global modeling ability of the entire network.Experimental results show that the improved network model not only reduces the number of parameters,but also improves the denoising performance.